JioHotstar compared six ways of giving a coding agent access to a codebase, using SWE-bench subsets, CodeScaleBench and an internal question harness, with Claude Sonnet solving. Plain filesystem tools stayed competitive on concrete tasks, Sourcebot's structural search did better on harder SWE-bench Verified tasks at lower cost, and semantic retrieval helped most when requirements were ambiguous. InfoQ's report gives no figures, so teams would need to check the original study before relying on the detail.
